MarcoForte / DeepInteractiveSegmentation

Getting to 99% Accuracy in Interactive Segmentation and Interactive Training and Architecture for Deep Object Selection

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Deep Interactive Segmentation

Official repository for the two papers.

Getting to 99% Accuracy in Interactive Segmentation, submitted to Signal Processing: Image Communication the Special Issue on Computational Image Editing.

Interactive Training and Architecture for Deep Object Selection, (Runner-up Best Paper) ICME 2020.
Marco Forte1, Brian Price2, Scott Cohen2, Ning Xu2, François Pitié1
1 Trinity College Dublin 2 Adobe Research

Requirements

GPU memory >= 4GB for inference on Berkeley and GrabCut. Optimal performance around 480p resolution.

Packages:

Additional Packages for jupyter notebook

  • matplotlib

Models

Model Name File Size NoC Grabcut NoC Berkeley
SyntheticPretrained+Finetune on SBD 144mb 1.74 2.93

Prediction

We provide a script demo.py which evaluates our model in terms of mean IoU and number of clicks to reach 90% accuracy. Links to download: the GrabCut and Berkeley datasets. Results are slightly improved from Table. 8 in the paper, this is due to changes in prediction, the weights are the same as used in the paper.

Training

Training code is not released at this time. It may be released upon acceptance of the paper.

Citation

@misc{forte2020InterSeg,
    title={Getting to 99% Accuracy in Interactive Segmentation},
    author={Marco Forte and Brian Price and Scott Cohen and Ning Xu and François Pitié},
    year={2020},
    eprint={2003.07932},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

Related works of ours

About

Getting to 99% Accuracy in Interactive Segmentation and Interactive Training and Architecture for Deep Object Selection

License:MIT License


Languages

Language:Jupyter Notebook 95.8%Language:Python 4.2%